SystemsSec 2018W Lecture 19: Difference between revisions
Pasted notes with no formatting |
Formatted the notes |
||
| Line 7: | Line 7: | ||
Every security technology has strengths and weaknesses | Every security technology has strengths and weaknesses | ||
Various types of data input for security software includes | '''Various types of data input for security software includes:''' | ||
IP Packets | * IP Packets | ||
System Calls | * System Calls | ||
Log files | * Log files | ||
Emails | * Emails | ||
Os statistics (resource usage) | * Os statistics (resource usage) | ||
Http traffic | * Http traffic | ||
etc | * etc | ||
The representation of the data is more important than the data itself | '''The representation of the data is more important than the data itself.''' | ||
Data representation is a machine learning concept. | * Data representation is a machine learning concept. | ||
Certain type of operations are easier to do on different data representations | * Certain type of operations are easier to do on different data representations | ||
Data can be converted into different types of representations | * Data can be converted into different types of representations | ||
For example if your input data is IP packets, the types of operations you can do are different than if you are able to have email messages extracted from them. | * For example if your input data is IP packets, the types of operations you can do are different than if you are able to have email messages extracted from them. | ||
Converting those IP packets to email messages is called a representation switch | ** Converting those IP packets to email messages is called a representation switch | ||
Data represented as emails would have the fields From, to, message, date sent, etc | ** Data represented as emails would have the fields From, to, message, date sent, etc | ||
Machine learning algorithms | * Machine learning algorithms | ||
Pattern recognition classification problem | ** Pattern recognition classification problem | ||
Inputs you want to label | ** Inputs you want to label | ||
Deep learning | === Deep learning === | ||
Deep learning can learn its own representation | * Deep learning can learn its own representation | ||
But it requires tons of data to train itself | * But it requires tons of data to train itself | ||
Deep learning takes a ton of time, because it has to process so much data | * Deep learning takes a ton of time, because it has to process so much data | ||
Most problems don’t have enough data for deep learning | * Most problems don’t have enough data for deep learning | ||
Adversarial Machine learning | === Adversarial Machine learning === | ||
Deep learning might lot learn the representation that you expect. | * Deep learning might lot learn the representation that you expect. | ||
While we think that it would learn that a stop sign is red, octagonal, contains the word STOP in white text, it could actually just be learning something trivial like what the tops of the letters look like. This would allow someone to deface a stop sign in some minor way such as putting dots above the letter that would make it so that the self driving cars can’t recognise the sign even though humans still could. | * While we think that it would learn that a stop sign is red, octagonal, contains the word STOP in white text, it could actually just be learning something trivial like what the tops of the letters look like. This would allow someone to deface a stop sign in some minor way such as putting dots above the letter that would make it so that the self driving cars can’t recognise the sign even though humans still could. | ||
| Line 47: | Line 46: | ||
For the example of a spam filter, it can have static policies for email addresses or keywords to ignore, and it can also learn from the vast amount of data available. Spam filtering is able to be quite sophisticated because there is a clear indication of success and failure, and there is incentive for both the host and the user to get it correct. | For the example of a spam filter, it can have static policies for email addresses or keywords to ignore, and it can also learn from the vast amount of data available. Spam filtering is able to be quite sophisticated because there is a clear indication of success and failure, and there is incentive for both the host and the user to get it correct. | ||
=== Log files: === | |||
* SIEMs (System information event management) | |||
** Manage logs from multiple systems | |||
** Represents all logs as a common representation for use with analytic tools | |||
** Not good in practice because you need domain specific knowledge to understand the logs | |||
*CASB | |||
** SIEM for cloud applications | |||
=== Graylisting === | |||
When a user firsts attempts to do something the server responds with responds with “try again later” instead of a failure, then after a set amount of time adds it to a list. When that thing is attempted again later the server accepts and adds that action to a white list because a spam bot will not try things twice, but a user will. | When a user firsts attempts to do something the server responds with responds with “try again later” instead of a failure, then after a set amount of time adds it to a list. When that thing is attempted again later the server accepts and adds that action to a white list because a spam bot will not try things twice, but a user will. | ||
Latest revision as of 01:37, 4 April 2018
Audio
Notes
Every security technology has strengths and weaknesses
Various types of data input for security software includes:
- IP Packets
- System Calls
- Log files
- Emails
- Os statistics (resource usage)
- Http traffic
- etc
The representation of the data is more important than the data itself.
- Data representation is a machine learning concept.
- Certain type of operations are easier to do on different data representations
- Data can be converted into different types of representations
- For example if your input data is IP packets, the types of operations you can do are different than if you are able to have email messages extracted from them.
- Converting those IP packets to email messages is called a representation switch
- Data represented as emails would have the fields From, to, message, date sent, etc
- Machine learning algorithms
- Pattern recognition classification problem
- Inputs you want to label
Deep learning
- Deep learning can learn its own representation
- But it requires tons of data to train itself
- Deep learning takes a ton of time, because it has to process so much data
- Most problems don’t have enough data for deep learning
Adversarial Machine learning
- Deep learning might lot learn the representation that you expect.
- While we think that it would learn that a stop sign is red, octagonal, contains the word STOP in white text, it could actually just be learning something trivial like what the tops of the letters look like. This would allow someone to deface a stop sign in some minor way such as putting dots above the letter that would make it so that the self driving cars can’t recognise the sign even though humans still could.
How the data is represented determines what tasks security technology can perform with it
Security technologies can apply static policies or learn patterns. Complicated systems will use both.
For the example of a spam filter, it can have static policies for email addresses or keywords to ignore, and it can also learn from the vast amount of data available. Spam filtering is able to be quite sophisticated because there is a clear indication of success and failure, and there is incentive for both the host and the user to get it correct.
Log files:
- SIEMs (System information event management)
- Manage logs from multiple systems
- Represents all logs as a common representation for use with analytic tools
- Not good in practice because you need domain specific knowledge to understand the logs
- CASB
- SIEM for cloud applications
Graylisting
When a user firsts attempts to do something the server responds with responds with “try again later” instead of a failure, then after a set amount of time adds it to a list. When that thing is attempted again later the server accepts and adds that action to a white list because a spam bot will not try things twice, but a user will.